INCV: An Inception Network Based Deep Learning Framework for Signature Recognition
Md. Hasibur Rahman, K. M. Aslam Uddin, Nusrat Jahan, Apurba Adhikary, Samrat Kumar Dey, Monishanker Halder, Avi Deb Raha, Mrityunjoy Gain, Sumit Kumar Dam, Yu Qiao, Anupam Kumar Bairagi · Khulna University Studies · 2024
The signature of an individual is a handwritten sign or mark that resembles an individual’s name, is often stylized and unique, and indicates the person’s identity, intent, and consent. There are many cases where the signature may be forged by an anonymous person, which is one of the most complicated real-world problems and has significant social and commercial impacts. Given the widespread use of handwritten signatures in legal and financial transactions, it is imperative for researchers to carefully choose an effective method to verify these signatures and prevent forgeries, which can result in significant financial losses for customers. Although there has been a lot of study done on forgery detection and signature verification, the difficulty of detecting competent forgeries remains a major problem for both scholars and practitioners. This paper developed a strategy called Inception Network Customized Version (INCV) for signature recognition based on the two latest inception network architectures: Inception Network v3 and Inception ResNet v2. We have collected the signature images of individuals and worked with these pre-trained models to apply transfer learning to create our customized models. We have employed our customized versions to recognize the images of the individual’s signature. Comparative analysis between the two customized versions of the inception network gives a better approach for recognizing individuals’ signatures than the traditional approaches for recognizing signatures, where INCV I (based on Inception Network V3) gives 97% accuracy on the train set and 92% accuracy on the test set, however, INCV II (based on Inception ResNet V2) produced 98% accuracy on the train set and 96% accuracy on the test set.